A word alignment study to improve the reliability of the statistical and neural translation system

Word alignment is an essential task for numerous natural language processing applications, including machine translation. The performance of the statistical machine translation systems is directly impacted by the performance of their alignment modules. However, such alignment models perform worse and induce low machine translation performance when translating morphological rich or low resource languages. The first objective of this paper is to examine the impact of incorporating some morphosyntactic features on the statistical alignment models and on the associated translation systems for the (Arabic, English) language pair, and to identify which of these features is most suitable. Although the neural machine translation system does not directly include a concept of word alignment, we propose, in the second part of this work, a method of adjusting the attention mechanism of these systems by the statistical alignments. Experimental results show that the proposed approaches significantly improve the alignment and the translation performances.